2 * threadpool-ms.c: Microsoft threadpool runtime support
5 * Ludovic Henry (ludovic.henry@xamarin.com)
7 * Copyright 2015 Xamarin, Inc (http://www.xamarin.com)
8 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
12 // Copyright (c) Microsoft. All rights reserved.
13 // Licensed under the MIT license. See LICENSE file in the project root for full license information.
16 // - src/vm/comthreadpool.cpp
17 // - src/vm/win32threadpoolcpp
18 // - src/vm/threadpoolrequest.cpp
19 // - src/vm/hillclimbing.cpp
21 // Ported from C++ to C and adjusted to Mono runtime
24 #define _USE_MATH_DEFINES // needed by MSVC to define math constants
29 #include <mono/metadata/class-internals.h>
30 #include <mono/metadata/exception.h>
31 #include <mono/metadata/gc-internals.h>
32 #include <mono/metadata/object.h>
33 #include <mono/metadata/object-internals.h>
34 #include <mono/metadata/threadpool-ms.h>
35 #include <mono/metadata/threadpool-ms-io.h>
36 #include <mono/metadata/w32event.h>
37 #include <mono/utils/atomic.h>
38 #include <mono/utils/mono-compiler.h>
39 #include <mono/utils/mono-complex.h>
40 #include <mono/utils/mono-lazy-init.h>
41 #include <mono/utils/mono-logger.h>
42 #include <mono/utils/mono-logger-internals.h>
43 #include <mono/utils/mono-proclib.h>
44 #include <mono/utils/mono-threads.h>
45 #include <mono/utils/mono-time.h>
46 #include <mono/utils/mono-rand.h>
47 #include <mono/io-layer/io-layer.h>
49 #define CPU_USAGE_LOW 80
50 #define CPU_USAGE_HIGH 95
52 #define MONITOR_INTERVAL 500 // ms
53 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
55 #define WORKER_CREATION_MAX_PER_SEC 10
57 /* The exponent to apply to the gain. 1.0 means to use linear gain,
58 * higher values will enhance large moves and damp small ones.
60 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
62 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
63 * of thread count, higher values bias more against higher thread counts.
65 #define HILL_CLIMBING_BIAS 0.15
67 #define HILL_CLIMBING_WAVE_PERIOD 4
68 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
69 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
70 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
71 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
72 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
73 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
74 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
75 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
76 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
77 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
81 gint16 max_working; /* determined by heuristic */
82 gint16 active; /* executing worker_thread */
83 gint16 working; /* actively executing worker_thread, not parked */
84 gint16 parked; /* parked */
91 /* Number of outstanding jobs */
92 gint32 outstanding_request;
93 /* Number of currently executing jobs */
95 /* Signalled when threadpool_jobs + outstanding_request is 0 */
96 /* Protected by threadpool->domains_lock */
97 MonoCoopCond cleanup_cond;
100 typedef MonoInternalThread ThreadPoolWorkingThread;
104 gint32 samples_to_measure;
105 gdouble target_throughput_ratio;
106 gdouble target_signal_to_noise_ratio;
107 gdouble max_change_per_second;
108 gdouble max_change_per_sample;
109 gint32 max_thread_wave_magnitude;
110 gint32 sample_interval_low;
111 gdouble thread_magnitude_multiplier;
112 gint32 sample_interval_high;
113 gdouble throughput_error_smoothing_factor;
114 gdouble gain_exponent;
115 gdouble max_sample_error;
117 gdouble current_control_setting;
118 gint64 total_samples;
119 gint16 last_thread_count;
120 gdouble elapsed_since_last_change;
121 gdouble completions_since_last_change;
123 gdouble average_throughput_noise;
126 gdouble *thread_counts;
128 guint32 current_sample_interval;
129 gpointer random_interval_generator;
131 gint32 accumulated_completion_count;
132 gdouble accumulated_sample_duration;
133 } ThreadPoolHillClimbing;
136 ThreadPoolCounter counters;
138 GPtrArray *domains; // ThreadPoolDomain* []
139 MonoCoopMutex domains_lock;
141 GPtrArray *working_threads; // ThreadPoolWorkingThread* []
142 gint32 parked_threads_count;
143 MonoCoopCond parked_threads_cond;
144 MonoCoopMutex active_threads_lock; /* protect access to working_threads and parked_threads */
146 guint32 worker_creation_current_second;
147 guint32 worker_creation_current_count;
148 MonoCoopMutex worker_creation_lock;
150 gint32 heuristic_completions;
151 gint64 heuristic_sample_start;
152 gint64 heuristic_last_dequeue; // ms
153 gint64 heuristic_last_adjustment; // ms
154 gint64 heuristic_adjustment_interval; // ms
155 ThreadPoolHillClimbing heuristic_hill_climbing;
156 MonoCoopMutex heuristic_lock;
158 gint32 limit_worker_min;
159 gint32 limit_worker_max;
163 MonoCpuUsageState *cpu_usage_state;
166 /* suspended by the debugger */
172 TRANSITION_INITIALIZING,
173 TRANSITION_RANDOM_MOVE,
174 TRANSITION_CLIMBING_MOVE,
175 TRANSITION_CHANGE_POINT,
176 TRANSITION_STABILIZING,
177 TRANSITION_STARVATION,
178 TRANSITION_THREAD_TIMED_OUT,
179 TRANSITION_UNDEFINED,
180 } ThreadPoolHeuristicStateTransition;
182 static mono_lazy_init_t status = MONO_LAZY_INIT_STATUS_NOT_INITIALIZED;
185 MONITOR_STATUS_REQUESTED,
186 MONITOR_STATUS_WAITING_FOR_REQUEST,
187 MONITOR_STATUS_NOT_RUNNING,
190 static gint32 monitor_status = MONITOR_STATUS_NOT_RUNNING;
192 static ThreadPool* threadpool;
194 #define COUNTER_CHECK(counter) \
196 g_assert (counter._.max_working > 0); \
197 g_assert (counter._.working >= 0); \
198 g_assert (counter._.active >= 0); \
201 #define COUNTER_READ() (InterlockedRead64 (&threadpool->counters.as_gint64))
203 #define COUNTER_ATOMIC(var,block) \
205 ThreadPoolCounter __old; \
207 g_assert (threadpool); \
208 __old.as_gint64 = COUNTER_READ (); \
211 COUNTER_CHECK (var); \
212 } while (InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
215 #define COUNTER_TRY_ATOMIC(res,var,block) \
217 ThreadPoolCounter __old; \
219 g_assert (threadpool); \
220 __old.as_gint64 = COUNTER_READ (); \
224 COUNTER_CHECK (var); \
225 (res) = InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) == __old.as_gint64; \
232 mono_coop_mutex_lock (&threadpool->domains_lock);
236 domains_unlock (void)
238 mono_coop_mutex_unlock (&threadpool->domains_lock);
245 return mono_rand_init (NULL, 0);
249 rand_next (gpointer *handle, guint32 min, guint32 max)
253 mono_rand_try_get_uint32 (handle, &val, min, max, &error);
254 // FIXME handle error
255 mono_error_assert_ok (&error);
260 rand_free (gpointer handle)
262 mono_rand_close (handle);
268 ThreadPoolHillClimbing *hc;
269 const char *threads_per_cpu_env;
270 gint threads_per_cpu;
273 g_assert (!threadpool);
274 threadpool = g_new0 (ThreadPool, 1);
275 g_assert (threadpool);
277 threadpool->domains = g_ptr_array_new ();
278 mono_coop_mutex_init (&threadpool->domains_lock);
280 threadpool->parked_threads_count = 0;
281 mono_coop_cond_init (&threadpool->parked_threads_cond);
282 threadpool->working_threads = g_ptr_array_new ();
283 mono_coop_mutex_init (&threadpool->active_threads_lock);
285 threadpool->worker_creation_current_second = -1;
286 mono_coop_mutex_init (&threadpool->worker_creation_lock);
288 threadpool->heuristic_adjustment_interval = 10;
289 mono_coop_mutex_init (&threadpool->heuristic_lock);
293 hc = &threadpool->heuristic_hill_climbing;
295 hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
296 hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
297 hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
298 hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
299 hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
300 hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
301 hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
302 hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
303 hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
304 hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
305 hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
306 hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
307 hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
308 hc->current_control_setting = 0;
309 hc->total_samples = 0;
310 hc->last_thread_count = 0;
311 hc->average_throughput_noise = 0;
312 hc->elapsed_since_last_change = 0;
313 hc->accumulated_completion_count = 0;
314 hc->accumulated_sample_duration = 0;
315 hc->samples = g_new0 (gdouble, hc->samples_to_measure);
316 hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
317 hc->random_interval_generator = rand_create ();
318 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
320 if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
323 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
325 threads_count = mono_cpu_count () * threads_per_cpu;
327 threadpool->limit_worker_min = threadpool->limit_io_min = threads_count;
329 #if defined (PLATFORM_ANDROID) || defined (HOST_IOS)
330 threadpool->limit_worker_max = threadpool->limit_io_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
332 threadpool->limit_worker_max = threadpool->limit_io_max = threads_count * 100;
335 threadpool->counters._.max_working = threadpool->limit_worker_min;
337 threadpool->cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
339 threadpool->suspended = FALSE;
342 static void worker_kill (ThreadPoolWorkingThread *thread);
349 /* we make the assumption along the code that we are
350 * cleaning up only if the runtime is shutting down */
351 g_assert (mono_runtime_is_shutting_down ());
353 while (monitor_status != MONITOR_STATUS_NOT_RUNNING)
354 mono_thread_info_sleep (1, NULL);
356 mono_coop_mutex_lock (&threadpool->active_threads_lock);
358 /* stop all threadpool->working_threads */
359 for (i = 0; i < threadpool->working_threads->len; ++i)
360 worker_kill ((ThreadPoolWorkingThread*) g_ptr_array_index (threadpool->working_threads, i));
362 /* unpark all threadpool->parked_threads */
363 mono_coop_cond_broadcast (&threadpool->parked_threads_cond);
365 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
369 mono_threadpool_ms_enqueue_work_item (MonoDomain *domain, MonoObject *work_item, MonoError *error)
371 static MonoClass *threadpool_class = NULL;
372 static MonoMethod *unsafe_queue_custom_work_item_method = NULL;
373 MonoDomain *current_domain;
377 mono_error_init (error);
378 g_assert (work_item);
380 if (!threadpool_class)
381 threadpool_class = mono_class_load_from_name (mono_defaults.corlib, "System.Threading", "ThreadPool");
383 if (!unsafe_queue_custom_work_item_method)
384 unsafe_queue_custom_work_item_method = mono_class_get_method_from_name (threadpool_class, "UnsafeQueueCustomWorkItem", 2);
385 g_assert (unsafe_queue_custom_work_item_method);
389 args [0] = (gpointer) work_item;
390 args [1] = (gpointer) &f;
392 current_domain = mono_domain_get ();
393 if (current_domain == domain) {
394 mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, error);
395 return_val_if_nok (error, FALSE);
397 mono_thread_push_appdomain_ref (domain);
398 if (mono_domain_set (domain, FALSE)) {
399 mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, error);
400 if (!is_ok (error)) {
401 mono_thread_pop_appdomain_ref ();
404 mono_domain_set (current_domain, TRUE);
406 mono_thread_pop_appdomain_ref ();
411 /* LOCKING: domains_lock must be held */
413 tpdomain_add (ThreadPoolDomain *tpdomain)
419 len = threadpool->domains->len;
420 for (i = 0; i < len; ++i) {
421 if (g_ptr_array_index (threadpool->domains, i) == tpdomain)
426 g_ptr_array_add (threadpool->domains, tpdomain);
429 /* LOCKING: domains_lock must be held. */
431 tpdomain_remove (ThreadPoolDomain *tpdomain)
434 return g_ptr_array_remove (threadpool->domains, tpdomain);
437 /* LOCKING: domains_lock must be held */
438 static ThreadPoolDomain *
439 tpdomain_get (MonoDomain *domain, gboolean create)
442 ThreadPoolDomain *tpdomain;
446 for (i = 0; i < threadpool->domains->len; ++i) {
447 ThreadPoolDomain *tpdomain;
449 tpdomain = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
450 if (tpdomain->domain == domain)
457 tpdomain = g_new0 (ThreadPoolDomain, 1);
458 tpdomain->domain = domain;
459 mono_coop_cond_init (&tpdomain->cleanup_cond);
461 tpdomain_add (tpdomain);
467 tpdomain_free (ThreadPoolDomain *tpdomain)
472 /* LOCKING: domains_lock must be held */
474 domain_any_has_request (void)
478 for (i = 0; i < threadpool->domains->len; ++i) {
479 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
480 if (tmp->outstanding_request > 0)
487 /* LOCKING: domains_lock must be held */
488 static ThreadPoolDomain *
489 tpdomain_get_next (ThreadPoolDomain *current)
491 ThreadPoolDomain *tpdomain = NULL;
494 len = threadpool->domains->len;
496 guint i, current_idx = -1;
498 for (i = 0; i < len; ++i) {
499 if (current == g_ptr_array_index (threadpool->domains, i)) {
504 g_assert (current_idx != (guint)-1);
506 for (i = current_idx + 1; i < len + current_idx + 1; ++i) {
507 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i % len);
508 if (tmp->outstanding_request > 0) {
519 worker_wait_interrupt (gpointer data)
521 mono_coop_mutex_lock (&threadpool->active_threads_lock);
522 mono_coop_cond_signal (&threadpool->parked_threads_cond);
523 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
526 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
530 gboolean timeout = FALSE;
532 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker parking", mono_native_thread_id_get ());
534 mono_gc_set_skip_thread (TRUE);
536 mono_coop_mutex_lock (&threadpool->active_threads_lock);
538 if (!mono_runtime_is_shutting_down ()) {
539 static gpointer rand_handle = NULL;
540 MonoInternalThread *thread_internal;
541 gboolean interrupted = FALSE;
544 rand_handle = rand_create ();
545 g_assert (rand_handle);
547 thread_internal = mono_thread_internal_current ();
548 g_assert (thread_internal);
550 threadpool->parked_threads_count += 1;
551 g_ptr_array_remove_fast (threadpool->working_threads, thread_internal);
553 mono_thread_info_install_interrupt (worker_wait_interrupt, NULL, &interrupted);
557 if (mono_coop_cond_timedwait (&threadpool->parked_threads_cond, &threadpool->active_threads_lock, rand_next (&rand_handle, 5 * 1000, 60 * 1000)) != 0)
560 mono_thread_info_uninstall_interrupt (&interrupted);
563 g_ptr_array_add (threadpool->working_threads, thread_internal);
564 threadpool->parked_threads_count -= 1;
567 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
569 mono_gc_set_skip_thread (FALSE);
571 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
577 worker_try_unpark (void)
579 gboolean res = FALSE;
581 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
583 mono_coop_mutex_lock (&threadpool->active_threads_lock);
584 if (threadpool->parked_threads_count > 0) {
585 mono_coop_cond_signal (&threadpool->parked_threads_cond);
588 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
590 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
596 worker_kill (ThreadPoolWorkingThread *thread)
598 if (thread == mono_thread_internal_current ())
601 mono_thread_internal_abort ((MonoInternalThread*) thread);
605 worker_thread (gpointer data)
608 MonoInternalThread *thread;
609 ThreadPoolDomain *tpdomain, *previous_tpdomain;
610 ThreadPoolCounter counter;
611 gboolean retire = FALSE;
613 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
615 g_assert (threadpool);
617 thread = mono_thread_internal_current ();
620 mono_thread_set_name_internal (thread, mono_string_new (mono_get_root_domain (), "Threadpool worker"), FALSE, &error);
621 mono_error_assert_ok (&error);
623 mono_coop_mutex_lock (&threadpool->active_threads_lock);
624 g_ptr_array_add (threadpool->working_threads, thread);
625 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
627 previous_tpdomain = NULL;
631 while (!mono_runtime_is_shutting_down ()) {
634 if ((thread->state & (ThreadState_AbortRequested | ThreadState_SuspendRequested)) != 0) {
636 mono_thread_interruption_checkpoint ();
640 if (retire || !(tpdomain = tpdomain_get_next (previous_tpdomain))) {
643 COUNTER_ATOMIC (counter, {
644 counter._.working --;
649 timeout = worker_park ();
652 COUNTER_ATOMIC (counter, {
653 counter._.working ++;
663 /* The tpdomain->domain might have unloaded, while this thread was parked */
664 previous_tpdomain = NULL;
669 tpdomain->outstanding_request --;
670 g_assert (tpdomain->outstanding_request >= 0);
672 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker running in domain %p (outstanding requests %d) ",
673 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
675 g_assert (tpdomain->domain);
676 g_assert (tpdomain->threadpool_jobs >= 0);
677 tpdomain->threadpool_jobs ++;
680 * This is needed so there is always an lmf frame in the runtime invoke call below,
681 * so ThreadAbortExceptions are caught even if the thread is in native code.
683 mono_defaults.threadpool_perform_wait_callback_method->save_lmf = TRUE;
687 mono_thread_push_appdomain_ref (tpdomain->domain);
688 if (mono_domain_set (tpdomain->domain, FALSE)) {
689 MonoObject *exc = NULL, *res;
691 res = mono_runtime_try_invoke (mono_defaults.threadpool_perform_wait_callback_method, NULL, NULL, &exc, &error);
692 if (exc || !mono_error_ok(&error)) {
694 exc = (MonoObject *) mono_error_convert_to_exception (&error);
696 mono_error_cleanup (&error);
697 mono_thread_internal_unhandled_exception (exc);
698 } else if (res && *(MonoBoolean*) mono_object_unbox (res) == FALSE)
701 mono_thread_clr_state (thread, (MonoThreadState)~ThreadState_Background);
702 if (!mono_thread_test_state (thread , ThreadState_Background))
703 ves_icall_System_Threading_Thread_SetState (thread, ThreadState_Background);
705 mono_domain_set (mono_get_root_domain (), TRUE);
707 mono_thread_pop_appdomain_ref ();
711 tpdomain->threadpool_jobs --;
712 g_assert (tpdomain->threadpool_jobs >= 0);
714 if (tpdomain->outstanding_request + tpdomain->threadpool_jobs == 0 && mono_domain_is_unloading (tpdomain->domain)) {
717 removed = tpdomain_remove (tpdomain);
720 mono_coop_cond_signal (&tpdomain->cleanup_cond);
724 previous_tpdomain = tpdomain;
729 mono_coop_mutex_lock (&threadpool->active_threads_lock);
730 g_ptr_array_remove_fast (threadpool->working_threads, thread);
731 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
733 COUNTER_ATOMIC (counter, {
738 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
742 worker_try_create (void)
744 ThreadPoolCounter counter;
745 MonoInternalThread *thread;
746 gint64 current_ticks;
749 mono_coop_mutex_lock (&threadpool->worker_creation_lock);
751 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
752 current_ticks = mono_100ns_ticks ();
753 now = current_ticks / (10 * 1000 * 1000);
754 if (0 == current_ticks) {
755 g_warning ("failed to get 100ns ticks");
757 if (threadpool->worker_creation_current_second != now) {
758 threadpool->worker_creation_current_second = now;
759 threadpool->worker_creation_current_count = 0;
761 g_assert (threadpool->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
762 if (threadpool->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
763 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of worker created per second reached, current count = %d",
764 mono_native_thread_id_get (), threadpool->worker_creation_current_count);
765 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
771 COUNTER_ATOMIC (counter, {
772 if (counter._.working >= counter._.max_working) {
773 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
774 mono_native_thread_id_get ());
775 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
778 counter._.working ++;
783 if ((thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, TRUE, 0, &error)) != NULL) {
784 threadpool->worker_creation_current_count += 1;
786 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, created %p, now = %d count = %d", mono_native_thread_id_get (), GUINT_TO_POINTER(thread->tid), now, threadpool->worker_creation_current_count);
787 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
791 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread due to %s", mono_native_thread_id_get (), mono_error_get_message (&error));
792 mono_error_cleanup (&error);
794 COUNTER_ATOMIC (counter, {
795 counter._.working --;
799 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
803 static void monitor_ensure_running (void);
806 worker_request (MonoDomain *domain)
808 ThreadPoolDomain *tpdomain;
811 g_assert (threadpool);
813 if (mono_runtime_is_shutting_down ())
818 /* synchronize check with worker_thread */
819 if (mono_domain_is_unloading (domain)) {
824 tpdomain = tpdomain_get (domain, TRUE);
826 tpdomain->outstanding_request ++;
828 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, domain = %p, outstanding_request = %d",
829 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
833 if (threadpool->suspended)
836 monitor_ensure_running ();
838 if (worker_try_unpark ()) {
839 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
843 if (worker_try_create ()) {
844 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
848 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
853 monitor_should_keep_running (void)
855 static gint64 last_should_keep_running = -1;
857 g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
859 if (InterlockedExchange (&monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
860 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
862 if (mono_runtime_is_shutting_down ()) {
863 should_keep_running = FALSE;
866 if (!domain_any_has_request ())
867 should_keep_running = FALSE;
870 if (!should_keep_running) {
871 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
872 should_keep_running = force_should_keep_running = TRUE;
877 if (should_keep_running) {
878 if (last_should_keep_running == -1 || !force_should_keep_running)
879 last_should_keep_running = mono_100ns_ticks ();
881 last_should_keep_running = -1;
882 if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
887 g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
893 monitor_sufficient_delay_since_last_dequeue (void)
897 g_assert (threadpool);
899 if (threadpool->cpu_usage < CPU_USAGE_LOW) {
900 threshold = MONITOR_INTERVAL;
902 ThreadPoolCounter counter;
903 counter.as_gint64 = COUNTER_READ();
904 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
907 return mono_msec_ticks () >= threadpool->heuristic_last_dequeue + threshold;
910 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
913 monitor_thread (void)
915 MonoInternalThread *current_thread = mono_thread_internal_current ();
918 mono_cpu_usage (threadpool->cpu_usage_state);
920 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
923 ThreadPoolCounter counter;
924 gboolean limit_worker_max_reached;
925 gint32 interval_left = MONITOR_INTERVAL;
926 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
928 g_assert (monitor_status != MONITOR_STATUS_NOT_RUNNING);
930 mono_gc_set_skip_thread (TRUE);
934 gboolean alerted = FALSE;
936 if (mono_runtime_is_shutting_down ())
939 ts = mono_msec_ticks ();
940 if (mono_thread_info_sleep (interval_left, &alerted) == 0)
942 interval_left -= mono_msec_ticks () - ts;
944 mono_gc_set_skip_thread (FALSE);
945 if ((current_thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0)
946 mono_thread_interruption_checkpoint ();
947 mono_gc_set_skip_thread (TRUE);
948 } while (interval_left > 0 && ++awake < 10);
950 mono_gc_set_skip_thread (FALSE);
952 if (threadpool->suspended)
955 if (mono_runtime_is_shutting_down ())
959 if (!domain_any_has_request ()) {
965 threadpool->cpu_usage = mono_cpu_usage (threadpool->cpu_usage_state);
967 if (!monitor_sufficient_delay_since_last_dequeue ())
970 limit_worker_max_reached = FALSE;
972 COUNTER_ATOMIC (counter, {
973 if (counter._.max_working >= threadpool->limit_worker_max) {
974 limit_worker_max_reached = TRUE;
977 counter._.max_working ++;
980 if (limit_worker_max_reached)
983 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
985 for (i = 0; i < 5; ++i) {
986 if (mono_runtime_is_shutting_down ())
989 if (worker_try_unpark ()) {
990 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
994 if (worker_try_create ()) {
995 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
999 } while (monitor_should_keep_running ());
1001 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
1005 monitor_ensure_running (void)
1009 switch (monitor_status) {
1010 case MONITOR_STATUS_REQUESTED:
1012 case MONITOR_STATUS_WAITING_FOR_REQUEST:
1013 InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
1015 case MONITOR_STATUS_NOT_RUNNING:
1016 if (mono_runtime_is_shutting_down ())
1018 if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
1019 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, TRUE, SMALL_STACK, &error)) {
1020 monitor_status = MONITOR_STATUS_NOT_RUNNING;
1021 mono_error_cleanup (&error);
1026 default: g_assert_not_reached ();
1032 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1034 ThreadPoolHillClimbing *hc;
1036 g_assert (threadpool);
1038 hc = &threadpool->heuristic_hill_climbing;
1040 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] hill climbing, change max number of threads %d", mono_native_thread_id_get (), new_thread_count);
1042 hc->last_thread_count = new_thread_count;
1043 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
1044 hc->elapsed_since_last_change = 0;
1045 hc->completions_since_last_change = 0;
1049 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1051 ThreadPoolHillClimbing *hc;
1053 g_assert (threadpool);
1055 hc = &threadpool->heuristic_hill_climbing;
1057 if (new_thread_count != hc->last_thread_count) {
1058 hc->current_control_setting += new_thread_count - hc->last_thread_count;
1059 hill_climbing_change_thread_count (new_thread_count, transition);
1063 static double_complex
1064 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
1066 ThreadPoolHillClimbing *hc;
1067 gdouble w, cosine, sine, coeff, q0, q1, q2;
1070 g_assert (threadpool);
1071 g_assert (sample_count >= period);
1072 g_assert (period >= 2);
1074 hc = &threadpool->heuristic_hill_climbing;
1076 w = 2.0 * M_PI / period;
1079 coeff = 2.0 * cosine;
1082 for (i = 0; i < sample_count; ++i) {
1083 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
1088 return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
1092 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, gint64 *adjustment_interval)
1094 ThreadPoolHillClimbing *hc;
1095 ThreadPoolHeuristicStateTransition transition;
1097 gdouble throughput_error_estimate;
1103 gint new_thread_wave_magnitude;
1104 gint new_thread_count;
1105 double_complex thread_wave_component;
1106 double_complex throughput_wave_component;
1107 double_complex ratio;
1109 g_assert (threadpool);
1110 g_assert (adjustment_interval);
1112 hc = &threadpool->heuristic_hill_climbing;
1114 /* If someone changed the thread count without telling us, update our records accordingly. */
1115 if (current_thread_count != hc->last_thread_count)
1116 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
1118 /* Update the cumulative stats for this thread count */
1119 hc->elapsed_since_last_change += sample_duration;
1120 hc->completions_since_last_change += completions;
1122 /* Add in any data we've already collected about this sample */
1123 sample_duration += hc->accumulated_sample_duration;
1124 completions += hc->accumulated_completion_count;
1126 /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
1127 * of each work item, we are goinng to be missing some data about what really happened during the
1128 * sample interval. The count produced by each thread includes an initial work item that may have
1129 * started well before the start of the interval, and each thread may have been running some new
1130 * work item for some time before the end of the interval, which did not yet get counted. So
1131 * our count is going to be off by +/- threadCount workitems.
1133 * The exception is that the thread that reported to us last time definitely wasn't running any work
1134 * at that time, and the thread that's reporting now definitely isn't running a work item now. So
1135 * we really only need to consider threadCount-1 threads.
1137 * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1139 * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1140 * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1141 * then the next one likely will be too. The one after that will include the sum of the completions
1142 * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1143 * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1144 * range we're targeting, which will not be filtered by the frequency-domain translation. */
1145 if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1146 /* Not accurate enough yet. Let's accumulate the data so
1147 * far, and tell the ThreadPool to collect a little more. */
1148 hc->accumulated_sample_duration = sample_duration;
1149 hc->accumulated_completion_count = completions;
1150 *adjustment_interval = 10;
1151 return current_thread_count;
1154 /* We've got enouugh data for our sample; reset our accumulators for next time. */
1155 hc->accumulated_sample_duration = 0;
1156 hc->accumulated_completion_count = 0;
1158 /* Add the current thread count and throughput sample to our history. */
1159 throughput = ((gdouble) completions) / sample_duration;
1161 sample_index = hc->total_samples % hc->samples_to_measure;
1162 hc->samples [sample_index] = throughput;
1163 hc->thread_counts [sample_index] = current_thread_count;
1164 hc->total_samples ++;
1166 /* Set up defaults for our metrics. */
1167 thread_wave_component = mono_double_complex_make(0, 0);
1168 throughput_wave_component = mono_double_complex_make(0, 0);
1169 throughput_error_estimate = 0;
1170 ratio = mono_double_complex_make(0, 0);
1173 transition = TRANSITION_WARMUP;
1175 /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1176 * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1177 * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1178 sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1180 if (sample_count > hc->wave_period) {
1182 gdouble average_throughput;
1183 gdouble average_thread_count;
1184 gdouble sample_sum = 0;
1185 gdouble thread_sum = 0;
1187 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1188 for (i = 0; i < sample_count; ++i) {
1189 guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1190 sample_sum += hc->samples [j];
1191 thread_sum += hc->thread_counts [j];
1194 average_throughput = sample_sum / sample_count;
1195 average_thread_count = thread_sum / sample_count;
1197 if (average_throughput > 0 && average_thread_count > 0) {
1198 gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1200 /* Calculate the periods of the adjacent frequency bands we'll be using to
1201 * measure noise levels. We want the two adjacent Fourier frequency bands. */
1202 adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1203 adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1205 /* Get the the three different frequency components of the throughput (scaled by average
1206 * throughput). Our "error" estimate (the amount of noise that might be present in the
1207 * frequency band we're really interested in) is the average of the adjacent bands. */
1208 throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
1209 throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
1211 if (adjacent_period_2 <= sample_count) {
1212 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1213 hc->samples, sample_count, adjacent_period_2), average_throughput)));
1216 /* Do the same for the thread counts, so we have something to compare to. We don't
1217 * measure thread count noise, because there is none; these are exact measurements. */
1218 thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1220 /* Update our moving average of the throughput noise. We'll use this
1221 * later as feedback to determine the new size of the thread wave. */
1222 if (hc->average_throughput_noise == 0) {
1223 hc->average_throughput_noise = throughput_error_estimate;
1225 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1226 + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1229 if (cabs (thread_wave_component) > 0) {
1230 /* Adjust the throughput wave so it's centered around the target wave,
1231 * and then calculate the adjusted throughput/thread ratio. */
1232 ratio = mono_double_complex_div (mono_double_complex_sub (throughput_wave_component, mono_double_complex_scalar_mul(thread_wave_component, hc->target_throughput_ratio)), thread_wave_component);
1233 transition = TRANSITION_CLIMBING_MOVE;
1235 ratio = mono_double_complex_make (0, 0);
1236 transition = TRANSITION_STABILIZING;
1239 noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1240 if (noise_for_confidence > 0) {
1241 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1243 /* there is no noise! */
1249 /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1250 * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1251 * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1252 * backward (because this indicates that our changes are having the opposite of the intended effect).
1253 * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1254 * having a negative or positive effect on throughput. */
1255 move = creal (ratio);
1256 move = CLAMP (move, -1.0, 1.0);
1258 /* Apply our confidence multiplier. */
1259 move *= CLAMP (confidence, -1.0, 1.0);
1261 /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1262 * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1263 * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1264 gain = hc->max_change_per_second * sample_duration;
1265 move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1266 move = MIN (move, hc->max_change_per_sample);
1268 /* If the result was positive, and CPU is > 95%, refuse the move. */
1269 if (move > 0.0 && threadpool->cpu_usage > CPU_USAGE_HIGH)
1272 /* Apply the move to our control setting. */
1273 hc->current_control_setting += move;
1275 /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1276 * throughput error. This average starts at zero, so we'll start with a nice safe little wave at first. */
1277 new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1278 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1279 new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1281 /* Make sure our control setting is within the ThreadPool's limits. */
1282 hc->current_control_setting = CLAMP (hc->current_control_setting, threadpool->limit_worker_min, threadpool->limit_worker_max - new_thread_wave_magnitude);
1284 /* Calculate the new thread count (control setting + square wave). */
1285 new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1287 /* Make sure the new thread count doesn't exceed the ThreadPool's limits. */
1288 new_thread_count = CLAMP (new_thread_count, threadpool->limit_worker_min, threadpool->limit_worker_max);
1290 if (new_thread_count != current_thread_count)
1291 hill_climbing_change_thread_count (new_thread_count, transition);
1293 if (creal (ratio) < 0.0 && new_thread_count == threadpool->limit_worker_min)
1294 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1296 *adjustment_interval = hc->current_sample_interval;
1298 return new_thread_count;
1302 heuristic_notify_work_completed (void)
1304 g_assert (threadpool);
1306 InterlockedIncrement (&threadpool->heuristic_completions);
1307 threadpool->heuristic_last_dequeue = mono_msec_ticks ();
1311 heuristic_should_adjust (void)
1313 g_assert (threadpool);
1315 if (threadpool->heuristic_last_dequeue > threadpool->heuristic_last_adjustment + threadpool->heuristic_adjustment_interval) {
1316 ThreadPoolCounter counter;
1317 counter.as_gint64 = COUNTER_READ();
1318 if (counter._.working <= counter._.max_working)
1326 heuristic_adjust (void)
1328 g_assert (threadpool);
1330 if (mono_coop_mutex_trylock (&threadpool->heuristic_lock) == 0) {
1331 gint32 completions = InterlockedExchange (&threadpool->heuristic_completions, 0);
1332 gint64 sample_end = mono_msec_ticks ();
1333 gint64 sample_duration = sample_end - threadpool->heuristic_sample_start;
1335 if (sample_duration >= threadpool->heuristic_adjustment_interval / 2) {
1336 ThreadPoolCounter counter;
1337 gint16 new_thread_count;
1339 counter.as_gint64 = COUNTER_READ ();
1340 new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &threadpool->heuristic_adjustment_interval);
1342 COUNTER_ATOMIC (counter, { counter._.max_working = new_thread_count; });
1344 if (new_thread_count > counter._.max_working)
1345 worker_request (mono_domain_get ());
1347 threadpool->heuristic_sample_start = sample_end;
1348 threadpool->heuristic_last_adjustment = mono_msec_ticks ();
1351 mono_coop_mutex_unlock (&threadpool->heuristic_lock);
1356 mono_threadpool_ms_cleanup (void)
1358 #ifndef DISABLE_SOCKETS
1359 mono_threadpool_ms_io_cleanup ();
1361 mono_lazy_cleanup (&status, cleanup);
1365 mono_threadpool_ms_begin_invoke (MonoDomain *domain, MonoObject *target, MonoMethod *method, gpointer *params, MonoError *error)
1367 static MonoClass *async_call_klass = NULL;
1368 MonoMethodMessage *message;
1369 MonoAsyncResult *async_result;
1370 MonoAsyncCall *async_call;
1371 MonoDelegate *async_callback = NULL;
1372 MonoObject *state = NULL;
1374 if (!async_call_klass)
1375 async_call_klass = mono_class_load_from_name (mono_defaults.corlib, "System", "MonoAsyncCall");
1377 mono_lazy_initialize (&status, initialize);
1379 mono_error_init (error);
1381 message = mono_method_call_message_new (method, params, mono_get_delegate_invoke (method->klass), (params != NULL) ? (&async_callback) : NULL, (params != NULL) ? (&state) : NULL, error);
1382 return_val_if_nok (error, NULL);
1384 async_call = (MonoAsyncCall*) mono_object_new_checked (domain, async_call_klass, error);
1385 return_val_if_nok (error, NULL);
1387 MONO_OBJECT_SETREF (async_call, msg, message);
1388 MONO_OBJECT_SETREF (async_call, state, state);
1390 if (async_callback) {
1391 MONO_OBJECT_SETREF (async_call, cb_method, mono_get_delegate_invoke (((MonoObject*) async_callback)->vtable->klass));
1392 MONO_OBJECT_SETREF (async_call, cb_target, async_callback);
1395 async_result = mono_async_result_new (domain, NULL, async_call->state, NULL, (MonoObject*) async_call, error);
1396 return_val_if_nok (error, NULL);
1397 MONO_OBJECT_SETREF (async_result, async_delegate, target);
1399 mono_threadpool_ms_enqueue_work_item (domain, (MonoObject*) async_result, error);
1400 return_val_if_nok (error, NULL);
1402 return async_result;
1406 mono_threadpool_ms_end_invoke (MonoAsyncResult *ares, MonoArray **out_args, MonoObject **exc, MonoError *error)
1410 mono_error_init (error);
1412 g_assert (out_args);
1417 /* check if already finished */
1418 mono_monitor_enter ((MonoObject*) ares);
1420 if (ares->endinvoke_called) {
1421 mono_error_set_invalid_operation(error, "Delegate EndInvoke method called more than once");
1422 mono_monitor_exit ((MonoObject*) ares);
1426 ares->endinvoke_called = 1;
1428 /* wait until we are really finished */
1429 if (ares->completed) {
1430 mono_monitor_exit ((MonoObject *) ares);
1432 gpointer wait_event;
1434 wait_event = mono_wait_handle_get_handle ((MonoWaitHandle*) ares->handle);
1436 wait_event = mono_w32event_create (TRUE, FALSE);
1437 g_assert(wait_event);
1438 MonoWaitHandle *wait_handle = mono_wait_handle_new (mono_object_domain (ares), wait_event, error);
1439 if (!is_ok (error)) {
1440 CloseHandle (wait_event);
1443 MONO_OBJECT_SETREF (ares, handle, (MonoObject*) wait_handle);
1445 mono_monitor_exit ((MonoObject*) ares);
1448 WaitForSingleObjectEx (wait_event, INFINITE, TRUE);
1450 mono_w32handle_wait_one (wait_event, MONO_INFINITE_WAIT, TRUE);
1455 ac = (MonoAsyncCall*) ares->object_data;
1458 *exc = ac->msg->exc; /* FIXME: GC add write barrier */
1459 *out_args = ac->out_args;
1464 mono_threadpool_ms_remove_domain_jobs (MonoDomain *domain, int timeout)
1467 ThreadPoolDomain *tpdomain;
1471 g_assert (timeout >= -1);
1473 g_assert (mono_domain_is_unloading (domain));
1476 end = mono_msec_ticks () + timeout;
1478 #ifndef DISABLE_SOCKETS
1479 mono_threadpool_ms_io_remove_domain_jobs (domain);
1480 if (timeout != -1) {
1481 if (mono_msec_ticks () > end)
1487 * Wait for all threads which execute jobs in the domain to exit.
1488 * The is_unloading () check in worker_request () ensures that
1489 * no new jobs are added after we enter the lock below.
1491 mono_lazy_initialize (&status, initialize);
1494 tpdomain = tpdomain_get (domain, FALSE);
1502 while (tpdomain->outstanding_request + tpdomain->threadpool_jobs > 0) {
1503 if (timeout == -1) {
1504 mono_coop_cond_wait (&tpdomain->cleanup_cond, &threadpool->domains_lock);
1509 now = mono_msec_ticks();
1515 res = mono_coop_cond_timedwait (&tpdomain->cleanup_cond, &threadpool->domains_lock, end - now);
1523 /* Remove from the list the worker threads look at */
1524 tpdomain_remove (tpdomain);
1528 mono_coop_cond_destroy (&tpdomain->cleanup_cond);
1529 tpdomain_free (tpdomain);
1535 mono_threadpool_ms_suspend (void)
1538 threadpool->suspended = TRUE;
1542 mono_threadpool_ms_resume (void)
1545 threadpool->suspended = FALSE;
1549 ves_icall_System_Threading_ThreadPool_GetAvailableThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1551 ThreadPoolCounter counter;
1553 if (!worker_threads || !completion_port_threads)
1556 mono_lazy_initialize (&status, initialize);
1558 counter.as_gint64 = COUNTER_READ ();
1560 *worker_threads = MAX (0, threadpool->limit_worker_max - counter._.active);
1561 *completion_port_threads = threadpool->limit_io_max;
1565 ves_icall_System_Threading_ThreadPool_GetMinThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1567 if (!worker_threads || !completion_port_threads)
1570 mono_lazy_initialize (&status, initialize);
1572 *worker_threads = threadpool->limit_worker_min;
1573 *completion_port_threads = threadpool->limit_io_min;
1577 ves_icall_System_Threading_ThreadPool_GetMaxThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1579 if (!worker_threads || !completion_port_threads)
1582 mono_lazy_initialize (&status, initialize);
1584 *worker_threads = threadpool->limit_worker_max;
1585 *completion_port_threads = threadpool->limit_io_max;
1589 ves_icall_System_Threading_ThreadPool_SetMinThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1591 mono_lazy_initialize (&status, initialize);
1593 if (worker_threads <= 0 || worker_threads > threadpool->limit_worker_max)
1595 if (completion_port_threads <= 0 || completion_port_threads > threadpool->limit_io_max)
1598 threadpool->limit_worker_min = worker_threads;
1599 threadpool->limit_io_min = completion_port_threads;
1605 ves_icall_System_Threading_ThreadPool_SetMaxThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1607 gint cpu_count = mono_cpu_count ();
1609 mono_lazy_initialize (&status, initialize);
1611 if (worker_threads < threadpool->limit_worker_min || worker_threads < cpu_count)
1613 if (completion_port_threads < threadpool->limit_io_min || completion_port_threads < cpu_count)
1616 threadpool->limit_worker_max = worker_threads;
1617 threadpool->limit_io_max = completion_port_threads;
1623 ves_icall_System_Threading_ThreadPool_InitializeVMTp (MonoBoolean *enable_worker_tracking)
1625 if (enable_worker_tracking) {
1626 // TODO implement some kind of switch to have the possibily to use it
1627 *enable_worker_tracking = FALSE;
1630 mono_lazy_initialize (&status, initialize);
1634 ves_icall_System_Threading_ThreadPool_NotifyWorkItemComplete (void)
1636 ThreadPoolCounter counter;
1638 if (mono_domain_is_unloading (mono_domain_get ()) || mono_runtime_is_shutting_down ())
1641 heuristic_notify_work_completed ();
1643 if (heuristic_should_adjust ())
1644 heuristic_adjust ();
1646 counter.as_gint64 = COUNTER_READ ();
1647 return counter._.working <= counter._.max_working;
1651 ves_icall_System_Threading_ThreadPool_NotifyWorkItemProgressNative (void)
1653 heuristic_notify_work_completed ();
1655 if (heuristic_should_adjust ())
1656 heuristic_adjust ();
1660 ves_icall_System_Threading_ThreadPool_ReportThreadStatus (MonoBoolean is_working)
1664 mono_error_set_not_implemented (&error, "");
1665 mono_error_set_pending_exception (&error);
1669 ves_icall_System_Threading_ThreadPool_RequestWorkerThread (void)
1671 return worker_request (mono_domain_get ());
1674 MonoBoolean G_GNUC_UNUSED
1675 ves_icall_System_Threading_ThreadPool_PostQueuedCompletionStatus (MonoNativeOverlapped *native_overlapped)
1677 /* This copy the behavior of the current Mono implementation */
1679 mono_error_set_not_implemented (&error, "");
1680 mono_error_set_pending_exception (&error);
1684 MonoBoolean G_GNUC_UNUSED
1685 ves_icall_System_Threading_ThreadPool_BindIOCompletionCallbackNative (gpointer file_handle)
1687 /* This copy the behavior of the current Mono implementation */
1691 MonoBoolean G_GNUC_UNUSED
1692 ves_icall_System_Threading_ThreadPool_IsThreadPoolHosted (void)